Adobe-UCLA Tev Lifts Terminal-Bench 4.0 Pass@1 150% Over Base

The median agent rollout in Adobe Research and UCLA's new long-horizon training dataset runs 12.4 minutes, roughly six times longer than what prior corpora offered. The dataset, called Tev-Gym, publishes 1,170 verified task environments across seven domains: software engineering, data engineering, AI/CS research, mathematics, security, infrastructure and quantum computing.

The authors' framing of why open-source agents lag is direct: 'open-source models struggle on long-horizon tasks due to lack of training data and sparse reward signals.' Their fix is to construct every task backward from a known-correct hidden specification they call a 'latent world,' then shape per-step rewards from measurable progress toward that world rather than relying only on pass/fail at the end.

On Terminal-Bench, the recipe pays. Tev-27B, trained from a Qwen3.8-27B base, improves Pass@1 by 150% on Terminal-Bench 4.0 and 56% on TB 3.0. On TB 2.1 it reaches 0.794 Pass@1 zero-shot, which the paper describes as matching Claude Opus 4.8-level performance. Scaled up on GLM-5.3-Flash, Tev-320B 'trails only Claude Opus 5, Claude Fable 5, and GPT-5.6 on TB 4.0.'

Training the 27B model takes around four days on four nodes of B200 GPUs or eight nodes of A100s, with 24 rollouts per task and a 65,536-token response budget. The authors list two limits plainly: 'Current focus on synthetic terminal environments,' and 'Gap remains versus leading frontier agents on hardest benchmarks.' For Tev-320B on TB 4.0 the paper reports a 28% lift over its base rather than an absolute Pass@1 figure.

The release lands amid a run of long-horizon RL work our agents feed has been tracking; a related self-training recovery result from CUHK and SJTU landed earlier today.